作者：Lucy Chai Richard Tucker Zhengqi Li Phillip Isola Noah Snavely
尽管图像质量越来越逼真，但最近的3D图像生成模型通常在固定范围的3D体积上操作，相机运动有限。我们研究了无条件合成无界自然，实现任意大的相机运动，同时保持持久的3D世界模型的任务。我们的场景表示由一个可扩展的平面布局网格组成，该网格可以通过3D解码器和体积渲染从任意相机姿势进行渲染，以及全景天穹。基于这种表示，我们仅从单视图互联网照片中学习生成世界模型。我们的方法能够在三维景观中模拟长距离飞行，同时保持全局场景的一致性——例如，返回起点会产生相同的场景视图。我们的方法能够超越当前3D生成模型的固定界限进行基因外推，同时也支持持久的、与相机无关的世界表现，这与au形成了鲜明对比
Despite increasingly realistic image quality, recent 3D image generativemodels often operate on 3D volumes of fixed extent with limited camera motions.We investigate the task of unconditionally synthesizing unbounded naturescenes, enabling arbitrarily large camera motion while maintaining a persistent3D world model. Our scene representation consists of an extendable, planarscene layout grid, which can be rendered from arbitrary camera poses via a 3Ddecoder and volume rendering, and a panoramic skydome. Based on thisrepresentation, we learn a generative world model solely from single-viewinternet photos. Our method enables simulating long flights through 3Dlandscapes, while maintaining global scene consistency–for instance, returningto the starting point yields the same view of the scene. Our approach enablesscene extrapolation beyond the fixed bounds of current 3D generative models,while also supporting a persistent, camera-independent world representationthat stands in contrast to auto-regressive 3D prediction models. Our projectpage: https://chail.github.io/persistent-nature/.
论文链接：http://arxiv.org/pdf/2303.13515v1
更多计算机论文：http://cspaper.cn/